Machine learning approach could help avoid computing crunch

Since the first detection in 2015 of ripples caused by the collision of blackholes in spacetime, astronomers have used powerful computers to analyse information from the signals, including mass, spin, polarisation and inclination of orbit of the bodies involved in each event.

With advances in technology at facilities such as the Laser Interferometer Gravitational-Wave Observatory (LIGO) in the USA and VIRGO in Italy, enabling detection of weaker signals, along with the development of on-line detectors like KAGRA in Japan, current computing infrastructure is under increasing pressure to analyse each signal which can take days to complete.

Researchers from the University of Glasgow have now trained a new system Vltamin,(1) to recognise gravitational wave signals from binary black holes using a machine learning technique called a conditional variational autoencoder, or CVAE. By passing simulated information through the system about 10 million times, VItamin improved its ability to pick out signals and analyse 15 parameters until it was capable of providing accurate results in less than a second.

Lead author Hunter Gabbard, School of Physics and Astronomy said: “Gravitational wave astronomy has provided us with an entirely new way to listen to the universe and the pace of developments since the first detection in 2015 has been remarkable”. 

“As detector technology improves further still and new detectors start listening too, we expect to be picking up hundreds of signals a year in the near future. Harnessing the power of machine learning will be vital to helping us keep pace with detector developments and VItamin is an exciting development towards that goal. We’re keen to work closely with colleagues in our global collaboration to integrate VItamin into the standard toolkit for detecting and responding to gravitational wave signals.”

Departmental colleague and co-author Dr Chris Messenger added: “Events like the collisions of black holes are invisible to electromagnetic telescopes, which is why we didn’t have direct evidence of their existence until the first gravitational wave signal from a black hole merger was detected. 

“But events like the collision of two neutron stars do have a visible component. In 2017, we detected the first gravitational wave signal from a neutron star merger and we were able to help our collaborators in electromagnetic astronomy turn their telescopes to the point in the sky where they could see the afterglow of the event."   

Professor Roderick Murray-Smith,School of Computing and co-author said: “The scientific domain of gravitational wave astronomy was a new area for us and it gave us the opportunity to design new models, specifically tailored for this application, which brought understanding of the physics together with leading edge machine learning methods. The University of Glasgow-led QuantIC project, which funded our work, has been a great opportunity to bring machine learning together with science, especially physics.”

Dr Francesco Tonolini, also from the School of computing and co-author added: “In the past, these latent variable models have been typically developed and optimised to capture distributions of images, text and other common signals. However, the data and distributions encountered in the gravitational wave domain are really quite unique and required us to re-invent components of these models to achieve the success demonstrated."

The research was supported by funding from the Science and Technology Facilities Council (STFC), the Science and Technology Research Council, the European Cooperation in Science and Technology (COST) action, Amazon Research and the Engineering and Physical Sciences Research Council (EPSRC).

‘Bayesian parameter estimation using conditional variational autoencoders for gravitational-wave astronomy’, Nature Physics Dec 20, 2021.

More information online